PRODUCTION SCENARIO
Cobalt Routing classifies support intents on gpt-5.6-luna against a 900 ms p95 budget and will accept at most a one-point accuracy loss. Measured on 200 labelled cases: none 78% at 240 ms, low 96% at 820 ms, medium 96% at 2.4 s, high 96% at 4.1 s.
Which effort level should ship?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: low, the cheapest level whose accuracy still matches medium exactly
Effort trades latency against accuracy, and the defensible pick is the cheapest level that clears both stated bars. none is the fastest measured but gives up eighteen points against a one-point allowance, medium matches on accuracy yet misses the 900 ms budget by more than double, and high is slower still while buying no accuracy at all. low already matches medium at 96% and lands at 820 ms, inside both constraints.
Official Source: https://developers.openai.com/api/docs/guides/reasoning ↗